HiDream-O1 Reference Images

HiDream-O1 is the 8B pixel-space model from HiDream.ai that opened under MIT in May 2026 - no VAE, no separate text encoder, just a transformer that consumes images and prompts directly. This node is how you hand it reference images. The tooltip gives you the whole map in one sentence: 1 image means instruction edit, 2-10 images means multi-reference.

Inputs

• positive / negative - your text conditioning. In HiDream-O1's world, that conditioning comes from its own prompt-encode nodes; this node layers images on top.
• images - the reference images, an autogrow input: wire in one to ten of them. One image switches the model into instruction-edit mode (change this photo according to the prompt); two or more put it in multi-reference mode for subject/style-driven generation.

Outputs

• positive / negative - the same conditioning, now carrying the reference images attached to both sides.

How it works

The mechanism is about as direct as conditioning gets: the node takes the raw image tensors and appends them to both conditionings under reference_latents. It doesn't encode, doesn't resize, doesn't fuse - the model consumes them in context during sampling, which is the same in-context design as Flux Kontext. It's been in core since May 2026, landing with HiDream-O1 support, and it's the node that makes O1's identity-preservation and editing tricks work.

What actually matters

The 1-vs-many distinction isn't cosmetic - it changes what the model is doing. With one reference you're saying "edit this image"; with several you're saying "make something in the spirit of these." The community's take on O1 was telling: it's a very good editor, with style transfer people called insane, even where its raw generation didn't displace the general local default. This node is the lever for both of those strengths.

Gotchas

Autogrow inputs can be fiddly - the empty slots are just placeholders, but if you wire images into slots out of order (image_3 filled while image_1 is empty), the node reads them in order and you'll get a confusing shuffle. Keep references in the order you want them consumed. And because it's pixel-space with no VAE, resolution is more sensitive than you'd expect from latent models - a huge reference can push memory; resize to a sane working size before it lands here. Negative conditioning gets the same references, which is correct for the model's contrast behavior, but if you're chasing a specific failure mode, that's the first thing to experiment with.

One context flag: HiDream's open/closed split confused people at launch - the leaderboard results people saw were often the closed 200B Pro sibling, not this MIT 8B. Don't set your expectations by the benchmark images; the 8B is an editor first. Ships with ComfyUI core, no install.